acceptodds
Withdrawn from ICLR 2027

EC-SleepMamba: An EEG-Guided Cross-Modal Mamba Framework for Sleep Staging

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Automatic sleep staging is an important component of quantitative sleep assessment and provides fundamental support for sleep analysis. Despite recent advances, existing multimodal fusion-based approaches face two challenges: (i) The symmetric fusion strategy ignores the asymmetric clinical prior that EEG is the primary modality, while EOG and EMG serve as auxiliary signals, overlooking the inherent discrepancy in their information contributions. (ii) Jointly modeling intra-epoch dynamics with inter-epoch stage transitions across long-duration physiological signals remains computationally demanding. To address these issues, we propose EC-SleepMamba, an EEG-guided cross-modal Mamba framework that leverages modality complementarity and temporal context to improve staging robustness. Specifically, modality-specific dual-scale CNNs are first introduced to extract high-resolution local features. During sequence encoding, an EEG-guided interaction gate is proposed to condition the auxiliary Cross-Mamba branches on EEG-derived context, enabling directional primary-to-auxiliary information regulation during representation encoding. Finally, a global bi-Mamba module is adopted to capture dynamic transition patterns across epochs. Experiments on Sleep-EDF-39 and Sleep-EDF-153 show that EC-SleepMamba outperforms five competing methods and improves the discrimination of highly confusable stages, demonstrating its effectiveness for sleep staging.

Primary area:
learning on time series and dynamical systems
Keywords:
multimodal sleep staging, asymmetric multimodal learning, state space models, Mamba, cross-modal representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_cRNyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes EC-SleepMamba, an EEG-guided multimodal framework for automatic sleep staging. It combines dual-scale CNNs, EEG-conditioned cross-modal gating, and hierarchical bidirectional Mamba modules to capture intra-epoch and inter-epoch dependencies.

Critical strengths

N/A

Critical weaknesses
  1. The experiments are limited to Sleep-EDF-39 and Sleep-EDF-153, two subsets of the same underlying dataset. The evaluation lacks independent datasets, broader generalization studies, and computational efficiency comparisons. The overall experimental scope falls substantially short of ICLR standards.

  2. All five experimental baselines were published in or before 2022, with SleepTransformer (2022) being the most recent. More recent sleep staging approaches, particularly Mamba-based and multimodal methods, are not evaluated. Consequently, the reported improvements provide insufficient evidence of competitiveness against the current state of the art.

  3. The proposed framework mainly combines established components, including dual-scale CNNs, bidirectional Mamba blocks, and a relatively straightforward EEG-conditioned gating mechanism. These modifications are largely incremental, without introducing a sufficiently novel methodological contribution to justify acceptance at ICLR.

Questions for authors

See in Weaknesses

Reviewer_TepwOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EC-SleepMamba, an EEG-guided multimodal Mamba framework for sleep staging. It first uses Dual-Scale CNNs to extract multi-scale features from EEG, EOG, and EMG signals. Then, Local Bi-Mamba is used to capture intra-epoch temporal dependencies, while an EEG-guided Interaction Gate is introduced to regulate the Mamba gates of EOG and EMG using EEG features. Finally, Global Bi-Mamba is used to model inter-epoch sleep stage transitions.

Critical strengths
  1. The overall framework is clear and well-organized, combining multi-scale feature extraction with both intra-epoch and inter-epoch temporal modeling.

  2. The main experiments and ablation studies are relatively comprehensive and show the effectiveness of the proposed method.

Critical weaknesses
  1. The methodological novelty is limited. The main contribution is an EEG-guided gate that regulates the features of other modalities, while most of the other components are based on existing methods.

  2. The baseline comparisons are somewhat outdated. More recent methods, such as BiT-MamSleep, SleepWaveNet, and TSEDSleepNet, are not included.

Questions for authors

The paper emphasizes the linear complexity and computational efficiency of Mamba, but does not provide computational cost comparisons. Could the authors report training or inference time, GPU memory usage, and FLOPs compared with the baselines?

Reviewer_zYRcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a multimodal temporal modeling framework for automatic sleep staging. It treats EEG as the primary information source and introduces an EEG-guided interaction gate to modulate EOG and EMG feature encoding. Local Bi-Mamba and Global Bi-Mamba are employed to model physiological dynamics within each 30-second epoch and sleep-stage transitions across epochs, respectively. The authors evaluate the proposed framework through main experiments and ablation studies on sleep-staging datasets.

Critical strengths
  1. Asymmetric multimodal signal processing. Rather than simply concatenating features from EEG, EOG, and EMG, the paper treats EEG as the primary modality and EOG and EMG as auxiliary modalities, using an EEG-guided interaction gate to modulate the feature encoding of the auxiliary modalities. This design reflects consideration of the potential differences in the information contributions of different physiological signals to sleep staging.
  2. Hierarchical temporal modeling. The framework separately models within-epoch physiological dynamics and across-epoch sleep-stage transitions, avoiding unified modeling of long-duration, high-sampling-rate multimodal signals. Its overall architecture is clearly structured.
Critical weaknesses
  1. Limited methodological novelty. The main technical contribution is the integration of EEG-guided conditional gating into the Cross-Mamba encoding process for auxiliary modalities, rather than performing simple feature fusion after encoding. However, the fundamental distinction from existing conditional gating, cross-modal feature modulation, and gated state-space models remains unclear. Placing a gating mechanism inside Mamba alone does not establish substantial algorithmic novelty.

  2. Insufficient evidence for the mechanism of EEG-guided gating. The current experiments do not adequately demonstrate that the gate selectively preserves auxiliary information relevant to the current brain state. Overall performance improvements cannot distinguish the specific benefits of EEG guidance from those of generic conditional modulation. Gate visualization, sleep-stage-specific analysis, and interventions using shuffled EEG context would provide more direct evidence.

  3. Insufficient justification of computational efficiency. Although the paper emphasizes Mamba's linear complexity, this does not necessarily imply superior computational efficiency for the entire framework. Bidirectional scans, multiscale CNNs, multiple modality branches, and additional gating mechanisms all introduce computational overhead. The authors should provide computational efficiency metrics and compare them with competitive baselines.

  4. Limited scope of experimental comparisons. The paper compares only a limited number of baseline methods, lacks comparisons with recent methods, and relies primarily on the Sleep-EDF datasets, resulting in insufficient experimental validation.

  5. Insufficient details on experimental settings and data splitting. Evaluation results for sleep staging may differ between cross-subject and non-cross-subject settings. The authors should clarify the evaluation protocol adopted and provide detailed information on the data splitting strategy and other experimental settings.

  6. Ablation studies don't sufficiently isolate the contribution of the core component. Removing the EEG-guided gate reduces Macro-F1 by 1.9 percentage points, whereas removing the auxiliary Cross-Mamba branch results in a reduction of only 0.7 points. This provides preliminary evidence for the value of the gating mechanism. However, module removal also changes model capacity and architecture, leaving open the possibility that the improvement partly comes from additional learnable parameters or generic feature modulation. Parameter-matched generic gating, shuffled EEG-context gating, and conventional cross-modal attention should be included as targeted controls.

  7. Insufficient stage-specific interpretability analysis. The paper emphasizes the relationship between EEG-guided gating and sleep stages but provides limited direct evidence of stage-specific gating behavior. Visualizations and stage-wise statistics of gate values would help establish whether the learned behavior is consistent with the proposed physiological priors.

  8. Claims regarding performance improvements should be more carefully qualified. Although the model shows notable improvements in stages such as N1 and REM, these results do not establish stronger discrimination across all commonly confused sleep stages. The authors should characterize the scope of the improvements more precisely using class-wise metrics and confusion matrices.

Questions for authors
  1. Distinction from existing methods. What specific, non-trivial technical differences distinguish the EEG-guided interaction gate from existing conditional gating, cross-modal feature modulation, and gated state-space models? Does its performance advantage specifically depend on injecting EEG conditions into Cross-Mamba, rather than using generic conditional gating or feature fusion?

  2. Validation of EEG-guided modulation. Could you provide additional experiments to validate the effectiveness of the EEG-guided interaction gate and better clarify its contribution to the model's performance improvements?

  3. Practical computational efficiency. Could you provide further experimental results on computational efficiency to support the claims regarding the model's efficiency?

  4. Evaluation protocols and generalizability. Does the evaluation follow a cross-subject, cross-session, or within-subject protocol? Could you clarify the evaluation protocol, and further validate the model's generalizability using more competitive baseline methods and additional datasets?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.